Positioning An official alpha package that adapts TypeSafe System One models, including Jev, to the Microsoft Agent Framework Python chat client contract. It maps TypeSafe's structured decision API onto that contract, so Jev evaluates application state against explicit typed questions and returns probabilities and scores rather than ordinary chat text (official).
What it does For this connector, Agent Framework's response_format option is a TypeSafe Questions mapping; the connector forwards it as the SDK's questions argument and uses SystemOneResponse as the actual response model. Noul, Choice, and Score questions are supported, and default_questions on the client lets a fixed contract omit per-call configuration. Tool selection and supported arguments are converted into internal Choice and Noul questions that emit Agent Framework function calls; MCP tools are supported through Agent, which expands discovered functions into FunctionTool objects. TypeSafeChatClient is the recommended client, layering function invocation, middleware, and telemetry over RawTypeSafeChatClient, which is for composing a custom layer stack. Internally created clients read TYPESAFE_API_KEY, TYPESAFE_DEFAULT_MODEL, and TYPESAFE_BASE_URL, defaulting to the jev-latest model at api.typesafe.ai. Streaming, non-text message content, and generative settings such as temperature are rejected (official).
Characteristics The package is alpha, installed with pip install agent-framework-typesafe --pre, and Python-only; .NET is not supported. It is MIT licensed. PyPI lists version 1.0.0a261002, released 2026-10-02 (PyPI). A request supports at most 32 routable tools, 64 properties per tool, 64 enum members per argument, and 128 generated internal questions (official).
When to use Suits Microsoft Agent Framework Python agents that need typed decisions for routing, tool-call gating, or judging. It does not fit .NET, streaming output, or tasks that require generative text.